Lecture 2.5
A condition is the point at which authority meets fact. This lecture examines evaluability: which conditions a machine can test, which require judgement, and how the boundary is declared.
Learning objectives
After this lesson, the reader should understand:
Concept framework
Condition classification
Case study
How should an artifact handle a condition that requires reasonableness?
Take one clause: 'An employee is eligible for compensation if they were a current employee at the time of injury, the injury is confirmed by a licensed medical practitioner's report, the employee took reasonable care to avoid the injury, and, where the injury involves a workplace safety violation, the case is referred to the health and safety regulator.' Four conditions, one sentence, and each belongs to a different class. 'Current employee at the time of injury' is evaluable from structured facts: check the employment record against the injury date, done. 'Confirmed by a licensed medical practitioner's report' is evaluable from attested evidence; the artifact does not diagnose the injury itself, it verifies that a qualified third party attested to it. 'Took reasonable care to avoid the injury' is judgement-bound: no database resolves what reasonable care meant in this specific situation. 'Where the injury involves a workplace safety violation, refer to the regulator' requires a further authority; the artifact is not positioned to decide the case at all, only to recognise that it belongs to someone else. Treated as prose, all four conditions read the same: things to satisfy before a payout. Classified, they demand four different behaviours. The first two run unattended. The third halts and refers, with the evaluated context attached. The fourth recognises it is not the deciding body, and routes the case out of the artifact entirely rather than attempt a verdict it was never authorised to make. None of this makes the artifact better than a human at judging reasonableness. It makes the four different kinds of not-knowing visible and distinct, instead of collapsing them into one silent 'pending' status that hides which kind it actually is. The value is not that the machine decided — it is that the boundary between what a machine can test and what it cannot is now explicit, recorded, and auditable, instead of being absorbed silently by whoever happened to handle the case. The strongest objection is that refusing to proxy is a luxury. 'Reasonable care' could be approximated — no recorded safety-training breach, no prior warning on file — and doing so would let most claims run unattended rather than halting every one. Sometimes the proxy is exactly right. But a proxy is not a translation of the condition; it is a new rule that replaces it, and a new rule needs an author with the authority to make it. Scholarship on automated administration has documented what happens when engineers make that substitution quietly[1]: policy changes that were never announced, consulted on, or approved arrive embedded in code. If the institution decides that 'no recorded breach' is what reasonable care will mean for this scheme, it should issue that as a rule — at which point the condition is evaluable, by authority, and visible as a choice. Two limits remain. The classification is itself contestable: whether 'confirmed by a licensed practitioner' is evidence or judgement depends on whether the artifact must also assess what the report says. And moving a condition into the attested-evidence class does not remove trust; it relocates it to the attester, whose own authority to attest now needs representing[2]. Each class answers a different question about who is trusted, not whether anyone is.
Discussion questions
Discussion (0)
No responses yet. Be the first to answer one of the questions above.
Start the course to join the discussion.
Exercise
Classify the conditions in one rule.